An embedded real-time compensation system for quartz flexure accelerometers based on a lightweight neural network

Abstract Quartz flexure accelerometers (QFAs) are widely used in high-precision inertial navigation due to their superior linearity, resolution, and stability. However, their accuracy is fundamentally limited by stochastic noise processes, including velocity random Walk (VRW), acceleration random walk (AccRW) and bias instability (BI). To overcome these limitations, this study proposes a bottleneck residual convolutional neural network (BR CNN) for real-time stochastic error compensation in QFAs. The BR CNN employs a compact encoder–decoder architecture with residual connections, enabling efficient noise suppression while maintaining low computational latency for embedded implementation. The model was trained using public inertial datasets, where synthetic stochastic noise consistent with QFA characteristics was added to evaluate denoising performance. A self-developed prototype integrating the proposed model with real-time signal acquisition and processing hardware was used for experimental validation. Results show that the BR CNN effectively reduces random error parameters, achieving VRW = 9.10 μg/√ Hz, RRW = 0.292 μg/√ Hz, and BI = 1.431 μg, indicating substantial improvement over uncorrected data (VRW = 15.68 μg/√ Hz, RRW = 0.504 μg/√ Hz, BI = 2.467 μg). The model attains an inference latency of 667 μs, verifying its real-time feasibility on microcontrollers. Relative to other CNN-based methods, the BR CNN provides an effective balance between accuracy and inference latency while operating under a conservative filtering strategy that minimizes the risk of signal distortion.